Training
Adjusting a model's parameters so that it does its job better on a given dataset. For a language model, that usually means lowering next-token loss across the training set.
Training a counts-table bigram is just incrementing counters. Training a neural network is running gradient descent on millions to trillions of parameters. The underlying loop — measure how wrong you are, change something to be less wrong, repeat — is the same.
Companion explanation in Step by Token, chapter 6.
Where this term gets built
- ch. 0Before you start
- ch. 1The dumbest model that exists
- ch. 2Counting isn't enough
- ch. 3Train your own tokens
- ch. 4Giving meaning to words
- ch. 5A neuron that learns
- ch. 6Stacking layers
- ch. 7Gradient descent live
- ch. 8An attention head by hand
- ch. 9Multi-head and residuals
- ch. 10The full transformer block
- ch. 11Prepare a dataset
- ch. 12The minimum code
- ch. 13The training loop
- ch. 15Load real weights
- ch. 16Why your model talks badly
- ch. 17Give your model instructions
- ch. 18Fine-tuning with LoRA
- ch. 19Simple quantization
- ch. 20Talk to your model
- ch. 21Ship a useful one
- ch. 22Appendix · Backprop by hand
- ch. 23Appendix · RLHF and DPO
Continue